Advancing Operational Earthquake Monitoring at Local and Regional Scales with Machine Learning-Enhanced SeisComP Tools - as Demonstrated in Switzerland

Machine learning (ML) has seen widespread use in seismology recently, with a significant focus on earthquake monitoring. ML models are now available for phase picking, first motion polarity determination, etc. Implementing them in standard monitoring software (e.g. SeisComP) could significantly improve the automatic earthquake monitoring and save time for human analysts, whilst leveraging all the existing benefits of existing mature monitoring frameworks. An important first step for moving the ML models from research into production has been the Python package SeisBench (Woollam et al., 2022; DOI: 10.1785/0220210324), which allows users to benchmark and access ML models and datasets. The scdlpicker SeisComP module (Tillman et al., 2023; DOI: 10.5194/egusphere-egu23-10046) created an interface between SeisComP and the trained ML pickers in SeisBench to allow event-based re-picking (i.e., not real-time phase onset detection) as demonstrated using teleseismic earthquakes and the GEOFON network. Here, we build on top of the existing scdlpicker module to provide both P and S picks at local distances, and add pick uncertainty and P-pick first motion polarity. We demonstrate the performance of this extended module in routine earthquake monitoring at the Swiss Seismological Service (SED) and show the improvements over classical pickers currently in use. We show that the ML pickers improve the automatic monitoring in both the number and the quality of the picks, leading to better automatic locations and magnitudes. We show that the ML picker’s characteristic function provides a good proxy of the human analyst assigned pick uncertainty. Additionally, this extended SeisComP module provides the ML-determined first-motion polarity for each pick, fully characterizing the pick itself (pick time, pick uncertainty, first motion polarity) in the same way a manual analyst would do. This allows the adoption of streamlined workflows in which the automatic (i.e. ML) picks would only be reviewed (and in most cases accepted) rather than re-picked from scratch by the human analyst (as currently done at SED).